Machine learning-based modeling of response patterns to transcranial direct current stimulation in women with fibromyalgia.

BACKGROUND

Clinical heterogeneity in fibromyalgia (FM) results in highly variable therapeutic responses to transcranial direct current stimulation (tDCS). Identifying predictors and temporal response patterns is crucial to transform personalized neuromodulation strategies. This secondary analysis integrated supervised prediction, pain-trajectory modeling, and causal inference to characterize individual response profiles to anodal tDCS in women with FM.

METHODS

Secondary analysis of a randomized, controlled, triple-blind clinical trial including 35 women allocated to active anodal M1 tDCS (2 mA, 20 min, 10 sessions) or sham stimulation. Pain (VAS) was assessed at baseline, D10, D30, and D90; functional, fatigue, mood, and QoL measures at baseline and D30. Six supervised algorithms were trained toclassify responders (≥30% VAS reduction). Pain-trajectory clusters were derived using Dynamic Time Warping with k-means. A T-learner framework estimated Conditional Average Treatment Effects (CATE) to detect treatment-effect modifiers. The Research Ethics Committee of the Centro Universitário da Grande Dourados under registration number Certificado de Apresentação de Apreciação Ética approved this research:36444920.5.0000.5159. The study was registered in The Brazilian Registry of Clinical Trials with the identifier RBR-8wc8rjq.

RESULTS

Active tDCS elicited greater pain reduction than sham at D10, D30, and D90 (p < .001). Predictive models demonstrated stable discrimination (AUC 0.78-0.84). WHOQOL-Physical and WHOQOL-Psychological domains were the strongest positive predictors, while higher FIQ and antidepressant use lowered response probability. Three exclusive pain-trajectory phenotypes were identified-rapid responders, progressive responders, and non-responders-with distinct sustainability at D90. CATE analyses indicated higher treatment advantage among participants with higher psychological QoL, moderate baseline VAS (6-8), and no antidepressant use.

CONCLUSION

The integration of machine-learning models, longitudinal trajectory analysis, and causal estimation revealed heterogeneous response dynamics and specific clinical profiles with higher likelihood of therapeutic gain from tDCS. These findings may support refinement of personalized stimulation protocols, early monitoring strategies, and targeted candidate selection in FM.